Finance, Administration, Legal and Customer Support

Administrative, financial and customer service work is made of documents, rules and repeated questions, which looks like a perfect fit for AI. It is also where an error can cost money, break a legal obligation or reveal confidential information. In this lesson you will see which tasks AI can support, where it must stop, and why a qualified professional stays in the loop.

We continue the case studies with Julien, the accountant, and Camille, who also handles customer inquiries for her company. A common thread runs through this lesson: AI can help you understand and draft; it does not give accounting, tax or legal advice, and it does not replace professional validation.

Julien's month-end

At the end of each month, Julien's firm chases clients for documents, reconciles bank lines with invoices, answers client questions about deadlines and drafts short explanatory notes. These tasks consume a large part of his time. He considers how AI could help while remembering that the firm owes confidentiality to its clients and is professionally responsible for its work.

Reconciliation and checks: help, don't hand over

Matching bank lines with invoices, spotting duplicates or unusual amounts is, at heart, a data task. Use what you learned in chapter 3: anonymize or use approved tools, ask for methods and formulas, test on small samples and verify. Specific cautions:

• AI can suggest a matching rule or explain a discrepancy pattern; it must not be the only control.
• Amounts, dates and references must be verified against source documents.
• Accounting treatment, tax positions and valuation judgments belong to the qualified professional, using the applicable rules.

Drafting reminders and routine letters

Payment reminders, requests for missing documents, appointment confirmations and standard letters are excellent tasks: repetitive, textual, easy to review. A good practice is to build a ladder of reminders: first friendly, second firm, third formal, each with a slightly different tone. Rules for debt collection and consumer communications vary by country and by situation, so the wording of formal notices and threats of action should follow your organization's validated templates and, where necessary, legal advice.

A ready-to-use prompt: reminder ladder

Write three versions of a reminder to a client about [unpaid invoice / missing documents]:
1. First reminder: friendly, assumes an oversight.
2. Second reminder: polite but clear about the consequences described in our terms: [only what is really in your terms].
3. Third reminder: formal, factual, stating the next step: [e.g. a call to agree a payment plan].
Each under 100 words. Keep [placeholders] for the client's name, invoice number, amount and dates. Do not threaten legal action, mention penalties or cite laws unless I provided the text. Tone: respectful at all times.

Understanding a contract (not advising)

A frequent use is summarizing a contract or a long legal text. This can help you understand a document: its structure, the main obligations, the dates, the points to ask a professional about. It is not legal advice, and the stakes are high:

• The assistant may miss a clause, misread an exception or reflect a legal system different from yours.
• Contracts are confidential and may contain personal data; use only tools approved for that purpose.
• The summary is not binding and should not be shared with others as if it were the contract.

A ready-to-use prompt: understand a clause

I am not a lawyer and I want to understand the clause below, in order to prepare questions for a qualified professional. This is NOT a request for legal advice.
Explain in plain language: what the clause says, who must do what and when, any deadlines or amounts, any exception or condition, and terms that I should ask a professional to explain.
Then list 5 questions I should ask a lawyer about it. If the clause is ambiguous, say so rather than choosing an interpretation. Do not tell me whether to sign.
Clause (with names and amounts replaced by placeholders):
[clause]

Confidentiality and professional secrecy

Some professions (accountants, lawyers, doctors and others) are subject to professional secrecy or specific confidentiality duties, and every organization has duties towards its clients and partners. Pasting client information into a tool can be a breach even with good intentions. Before doing so, ask:

1. Am I allowed to share this information with an external provider at all?
2. Does the tool's contract and configuration meet our obligations (data location, retention, no reuse for training)?
3. Can I remove or replace identifying details and still get the help I need?

When the answer to the first two questions is unclear, the answer is to anonymize or not to use the tool.

Customer support: FAQs and reply assistants

Customer service is a major field of AI use. Camille's company receives the same fifty questions about delivery, returns and assembly. There are three levels of use, with increasing risk.

1. Internal helper: the assistant drafts answers from your approved FAQ and policies for a human agent to review and send. Low to medium risk.
2. FAQ and help pages: you use AI to write and update help content from your real policies, verified by someone who knows them. Low risk.
3. Customer-facing chatbot: customers talk directly to an automated assistant. Higher risk: wrong promises about refunds or warranties, mishandled complaints, personal data in conversations, and transparency expectations (customers should know they are interacting with an automated system; the EU AI Act contains transparency obligations in certain cases, so check the official texts). Always provide an easy path to a human, define what the assistant must not answer and test it extensively before launch.

A ready-to-use prompt: reply from the approved FAQ

You help a customer service agent draft replies. Use ONLY the information in the FAQ below. If the answer is not in the FAQ, reply exactly: "Needs a human answer", and explain what is missing.
Do not promise refunds, deadlines, compensation or exceptions. Keep a [friendly, clear] tone, under 120 words, with [First name] as a placeholder.
FAQ:
[paste your approved policies]
Customer question (without personal data):
[question]

Bad use versus good use

Bad use	Good use
Pasting a client's full tax file into a chatbot to "check it".	Describing the situation without identifiers, or using an approved tool and a qualified review.
Following an AI summary of a contract and signing.	Using the summary to prepare questions for a lawyer, who gives the advice.
Launching a customer chatbot with no human fallback and no testing.	Starting with an internal reply assistant, then testing carefully before any public use, with a clear route to a human.

Your turn: draw the line

Exercise. List six tasks from administration, finance, legal or customer support in your context (or Julien's). Classify each in three columns: "AI can draft, I review", "AI only with an approved tool and anonymized data", and "Professional only". Then write the reminder ladder prompt for one real situation, or the FAQ reply prompt using five real questions from your customers.

See a solution

An example for Julien: "AI can draft, I review": reminder emails, meeting-confirmation messages, a plain-language note on a deadline using a text he provides. "Approved tool and anonymized data": classifying bank line descriptions, summarizing a long regulatory text. "Professional only": choosing a tax treatment, advising a client on a dispute with the tax authority, signing off the accounts.

For Camille's FAQ prompt, two of five test questions return "Needs a human answer" because the FAQ does not cover them. This is a good result: it reveals gaps in the FAQ rather than hiding them with a plausible invention. She adds those two items to her to-do list and writes the missing policies, with her manager's approval.